Locating and Editing Factual Knowledge in Large Language Models, Kevin Meng, David Bau, Huiyuan Chen, Zhiting Hu, Ofir Press, Meng Qu, Alexander Rush, 2022Advances in Neural Information Processing Systems (NeurIPS), Vol. 35 (NeurIPS)DOI: 10.5591/978-1-7138-3094-1.921 - Introduces ROME, a foundational method for directly editing factual associations in LLMs by modifying specific weights, offering precision and efficiency.
Mass-Editing Memory in a Transformer, Kevin Meng, Arnab Ghosh, David Bau, Raphael Gontijo Lopes, Sharan Narang, Jonathan Frankle, Ofir Press, Alexander Rush, 2023International Conference on Learning Representations (ICLR)DOI: 10.48550/arXiv.2210.07289 - Extends ROME to enable simultaneous editing of multiple facts efficiently, addressing scalability challenges in model editing.